[11:59 Sun,7.May 2023 by Thomas Richter]
NVIDIA has presented a new algorithm that can simulate entire tennis matches. However, the new method works differently than classic 3D simulations, which have to learn the different movements of a sport by means of elaborate motion capturing. The new system, on the other hand, is only trained using a number of ordinary television recordings of tennis matches. Unlike motion capture data, these are available in large quantities for popular sports such as tennis or soccer - but in poorer quality, because body parts are often obscured or the motion resolution is too poor.
In order to use this rather rough analysis data from TV recordings for the exact animation of 3D characters - such as hitting a ball with a racket in tennis - some additional steps have to be taken first. For example, video clips are first used to estimate player motion, and then a physical model of the human musculoskeletal system is used to correct the data.
A neural network then uses this data to learn entire sequences of movements typical of the sport in question - such as tennis in this case. An incoming ball can thus be hit to different target points via correct racket guidance or These sequences are then "assembled" into meaningful actions by a special control algorithm, thus synthesizing two physically simulated characters playing arbitrary extended rallies including correct simulated racket and ball dynamics.
Using the learned data, the system can simulate tennis players who can accurately hit the incoming ball to the target positions with a variety of stroke variations such as serve, forehand and backhand, spins (topspins and slices) and playing styles (one/two-handed backhand, left/right-handed). It can also correctly simulate different player types such as left-handed, right-handed or ambidextrous.
more infos at bei research.nvidia.com
deutsche Version dieser Seite: Neue NVIDIA-KI kann Tennisspiele simulieren - nur anhand von Fernsehbildern